Enhancing digital twin technology with community-led,science-driven participatory modeling: A case in green infrastructure planning

Abstract

Recent research, professional, and funding agendas have re-surfaced the importance of knowledge co-production and ethical participation to address urban tensions worldwide: urbanization and rapid climate change, disproportionately impacting socially vulnerable populations. Despite the rise of Digital Twins (DT), buoyed by the growth of computational and data technologies in the past 10 to 15 years, DT have fallen short of their promise to address these tensions. We present a participatory modeling (PM) platform, Fora.ai, to build on existing strengths of DT and overcome the most prevalent limitations of data-driven technologies. This platform (i.e., a set of visualization and simulation tools and facilitation and sense-making approaches) is organized around the iterative steps in PM: problem definition and goal setting, preference elicitation, collaborative scenario-building, simulation, tradeoff deliberation, and solution-building. We demonstrate the platform’s effectiveness when set within a stakeholder-led process that integrates diverse knowledge, data sources, and values in pursuit of equitable green infrastructure (GI) planning to address flooding. The immediate visualization of simulated impacts, followed by reflection on causal and spatial relationships and tradeoffs across diverse priorities, enhanced participants’ collective understanding of how GI interacts with the built environment and physical conditions to inform their intervention scenarios. The facilitated use of Fora.ai enabled a collaborative socio-technical sense-making process, whereby participants transitioned from untested beliefs to designs that were specifically tailored to the problem in the study area and the diversity of values represented, attending to both localized flooding and neighborhood-level impacts. They also derived generalizable design principles that could be applied elsewhere. We show how the combination of specific facilitation practices and platform features leverage the power of data, computational modeling, and social complexity to contribute to collaborative learning and creative and equitable solution-building for urban sustainability and climate resilience.

Introduction

The impacts of climate change and environmental hazards are not distributed evenly across the globe, or even within a region. Scientific models of global environmental hazards reveal these inequities at a global scale, and have been somewhat downscaled to a regional scale, but very few models characterize local variation at a high resolution. Recent research has revealed the presence of microspatial inequities that arise at hyperlocal scales within communities, creating disparate experiences and outcomes for people who live even just around the corner from each other (O’Brien et al., 2020; O’Brien and Mueller, 2023). Mitigating these hazards may thus require very local and context-specific solutions. Modeling the variation in the built environment, specifically in the composition and configuration of different types of green infrastructure (GI), can reveal drivers and explain uneven impacts of the built environment on affected communities (Derkzen et al., 2015; Wang et al., 2015). Given finite space and resources, targeted GI investment that leverage public or private space in areas of highest impact has the potential of mitigating local inequities, if technology can effectively measure and model those differences. However, planners typically develop high-level guidelines and regulations, driven by their understanding of urban landscapes and broad goals that may not fully consider the specifics of each site and community. On the other hand, designers (e.g., architects, engineers, and landscape architects) work in the specifics of a site and may model local impacts in complex engineering models, in a process that is separate from a reactive community approval process, and therefore may not fully account for the effects of those decisions outside of a limited boundary or user group (Zellner, 2008).

The emergence of Digital Twins (DT) promised to support planning and policy, particularly in highly complex urban settings. By creating a mirror image of a place—be it a city neighborhood, a coastal harbor, a forest, or otherwise—and simulating the processes that shape local conditions, they make scientific insights inherently accessible for practical decision-making. Their power is further enhanced by the proliferation of highly detailed data in recent years, allowing us to observe and model disparities across space and time with unprecedented precision. While the technology marches forward, however, a crucial ingredient has been lacking: sense-making to inform decisions, assuming the automatic benefit of data (Yossef Ravid and Aharon-Gutman, 2023). The data are often collected just data for data’s sake, the mirror reflecting an image that is not interpreted. Data quality and availability, and the effort and skills required for effective and meaningful computation, make DT particularly vulnerable in its application to planning and policy (Ferré-Bigorra et al., 2022; Hämäläinen, 2021). DT and their underlying data may not be able to catalyze transformations in policy and practice on their own.

We reflect on this situation by reviewing the trajectory of planning support tools leading to the emergence of DT (see Supplemental Materials S1). We focus on how the evolution of this field has traditionally centered on data technologies but less on the way users interact with the technology or how it can enhance sense-making of complexity and solution-building. While urban simulations can be helpful in urban decision-making, the demand for accuracy of predictions and (3D) visualization distracts from the purpose of the modeling and critical thinking (Goldstone and Wilensky, 2008; Zellner et al., 2022). Representational realism is also tied to the possibility of automation to avoid the “interference” of human emotion, but this does not equate to better judgments; computational algorithms are vulnerable to hidden biases and lack human intuition, while human assumptions can be more openly inspected through modeling.

We also highlight the persistent challenges to incorporate social dimensions in both the technology and the processes that use it. DT has not been successful at promoting community engagement and participation, either (Abdeen and Sepasgozar, 2022; Adade and De Vries, 2023). We further examine the role that the search for accuracy, realism, and real-time data flows have had in the trajectory of decision-support systems, potentially steering them away from serving their intended purpose with costly and distracting detail. More realistic representations and accuracy may not lead to better decisions (Lempert et al., 2020) and may in fact obscure understanding of complex problems and their inherent uncertainty. It is this understanding of how a system operates, however, that allows us to anticipate possible trajectories and inform how interventions in the system may play out and why. Data-driven realism and predictive accuracy, in and of itself, does not generate that understanding. The emphasis on visual replication and universal use thus tends to hinder the use of DT technology for decision-making, relative to participatory modeling (PM) processes that focus on the co-construction of knowledge, models, and datasets, and building trusting relationships among people and with the technology. DT could benefit, in fact, from PM practices that support collaborative exploration of problems and solutions with adequate engagement, facilitation, and assessment support (e.g., Hedelin et al., 2021).

This review frames the design of our platform, Fora.ai, and its application to participatory GI planning, and environmental planning more broadly. Fora.ai is a set of software and modeling tools that are integral to a facilitated collaborative learning and deliberation process by which users can jointly understand the geophysical dynamics of environmental hazards in an urban neighborhood, and how specific interventions might be expected to alter the ability of the landscape to mitigate and adapt to that hazard. Like DT, the platform integrates urban data and science-based models, but also incorporates community knowledge and values to define the problem and design and test various solutions within the span of a short workshop. Humans, rather than the DT technology, act as connectors of all these elements through the facilitated collaborative practice. We build on a specific use case to show how Fora.ai can address some of the most common but persistent shortcomings of DT, with visualization and prioritization features, and a highly responsive and adaptive computational model of the built environment, within a facilitated PM process. This socio-technical sense-making setup is designed to make GI planning and design processes more meaningful, inclusive, and equitable.

In the following sections, we describe the Fora.ai platform and two applications in different contexts, one with local planners and staff and the other with a community of earth systems modelers to inform GI planning for flooding in Chelsea, a coastal environmental justice municipality of approximately 40,000 inhabitants in Massachusetts, USA. We follow with an analysis of the discussions emerging from the PM with Fora.ai as participants developed plans and tested them through simulations to derive design and planning insights. We finalize with implications for future development and application of these tools and associated practices to urban planning and policy.

A different platform: Fora.ai

Fora.ai is an online PM platform for groups to explore complex, real-world problems and their solutions. Fora.ai is an envelope meant to connect users with parsimonious computational models of specific neighborhoods and environmental processes, such as stormwater runoff and urban heat island effects, to support agile scenario design, exploration, and sense-making. A communication protocol linked the platform frontend (coded in HTML, CSS, and Javascript with ReactJS), backend (using NodeJS and MySQL, including a SQL database), and model, all running from the same server. This setup allows users to collectively design solutions to these problems and test them with the plugged-in model to provide rapid feedback, iterating multiple times within a short (2-h) workshop to enhance collaborative learning and problem-solving, and support democratic decision-making (Zellner 2024).

Fora.ai is intended for facilitated stakeholder workshops conducted after collectively defining a problem space and outcomes of interest, geographic focus, and the data and modeling tools to work with. Facilitators, researchers, and key stakeholders collectively select up to 10 relevant outcomes—a greater number would increase the cognitive load and hamper the sense-making process (Zellner 2024). Users can also revise model parameters and their values to better match local conditions represented in the model of choice, before prioritizing outcomes.

Workshops meet either virtually or in-person; participants need a laptop, tablet or mobile phone running iOS and laptops running Windows or MacOS, and an Internet connection. During a workshop, participants gather in small groups and interact with the platform and with each other to jointly simulate how starting conditions lead to spatiotemporal outcomes of concern, and collaboratively build solutions that improve over this baseline. The four main components of Fora.ai with which users interact are concern profiles, a map and whiteboard, simulation results, and a group overview.

The concern profile

Concerns describe diverse types of outcomes for the problem being modeled (e.g., property damage, area flooded). Diversity in how individuals value concerns is inherent to social complexity, and Fora.ai captures these differences through concern profiles (Zellner 2024).

To start, each participant logs into Fora.ai and creates their individual Concern Profile, where they rank outputs and assign each a numerical weight corresponding to its relative importance (see Figure 1); weights increase with increasing importance. The values matter only relative to the weights assigned to other concerns and thus are normalized between 0 and 1 once they are saved (see Supplemental Materials S4 for more details). Participants are free to populate their profiles, however, they choose based on their own personal views, their professional expertise, collective goals, or role-playing as an external stakeholder. Concerns can be adjusted at any time during a workshop, and the data is stored on the central server’s SQL database.

Figure 1.

Sample concern profile setup for urban flooding.

Map and whiteboard

The centerpiece of Fora.ai is a map showing a stylized representation of a real-world location in a two-dimensional lattice of cells (see Figure 2). The spatial resolution is high for an urban setting, normally set to around 10 m × 10 m. The base data visualization shows land cover and key features, like street names and tree canopy, read from the server, but other data layers can be included, depending on the problem, the model used, and available data. The map is zoomable, rotatable, and pannable, like in commonly used online mapping applications. Users can click on a cell to examine its properties or to leave a comment visible to everyone, perhaps to point out a location that needs attention.

Figure 2.

Whiteboard. On the map: yellow = buildings; light green = permeable surfaces, for example, lawns; dark green circles = trees; dark gray (with white labels) = roads; medium gray = alleys; light gray = other impervious surfaces, for example, walkways, sidewalks, parking lots; red squares, orange circles, and light green squares = GI installations). Toolbar: GI tokens, commenting, data visualization.

The map serves as a shared whiteboard for participants to collectively create solutions (in our use case, GI layouts) to improve outcomes (reduce flooding and its negative impacts). Participants in the session are shown on the shared whiteboard with a cursor attached to their name. The activity of each participant is thus visible to others in the same session. Each problem has its own unique set of intervention options, represented as tokens users click and place and/or delete on the map (representing GI types). Rules restrict where tokens can be placed (e.g., green roofs on buildings but not roads). Facilitators assist the participant group in systematically designing and exploring GI placement strategies (e.g., start by focusing on one type of GI type or one spatial strategy at a time). Users place the tokens individually while they strategize as a group, communicating directly and through the whiteboard, synchronized through the central server. Once a design is deemed finalized, it is named and a designated person in the group starts the simulation. The platform exports the scenario as an input to the plugged-in model, which runs the simulation.

Simulation model and visualization of results

In PM, simulations must run fast, one of the reasons for using parsimonious models. We used the Landscape Green Infrastructure Design (L-GrID) model developed in NetLogo 6.2.2 (Wilensky, 1999) to assist decision-makers in understanding neighborhood flooding and address it with GI (Zellner et al., 2016; Zellner and Massey, 2024). The benefits of mimicking natural systems to collect, treat, and infiltrate rain where it falls are often assumed but remained untested, ignoring important interactions with other hydrological aspects of the system, such as roads and sewers, and their cumulative impact on urban stormwater hydrology. L-GrID is a cellular, process-based model that allows for agile exploration of GI scenarios under different rainfall conditions and produces outputs that include estimates of the extent of the flooded area and runoff volume captured by sewers, GI, and downstream areas, as well as percentage capacity of GI used in each storm, present value of GI installation and maintenance over the lifetime of the infrastructure, efficiency in $/gallon of runoff captured by GI, and an estimate of property damages. In order of execution, the processes are: precipitation, infiltration, sewer intake and treatment, evapotranspiration, surface flow, and watershed export downstream. Embedding L-GrID in Fora.ai allows users to interact in collaborative scenario building, modeling, sense-making, and trade-off deliberation around GI solutions to urban flooding. The model was parameterized to reflect Chelsea conditions (see Supplemental Materials S2 for more details on the L-GrID model, its inputs, and its output metrics).

Pressing the simulation button creates a file on the server with a list of the GI locations in a scenario, formats it for L-GrID input, and moves it to the L-GrID directory. The platform executes a command line to start the simulation, which takes less than 3 minutes to run for each GI scenario in the geographic scope selected (representing about 0.4 km²). L-GrID exports the spatial data at defined intervals during the simulation, and a collection of files with results aggregated across the whole landscape.

Once completed, the simulation is shown on playback video with the dynamics of a key output (ponded water depths) overlaid on the default view (land cover) (Figure 3).

Figure 3.

Simulation snapshot of shallow ponded water (blue shades) and flooding (pink).

A results tab shows the performance of each scenario across the concerns (Figure 4). The default view shows the aggregated result for each concern (e.g., the percentage total precipitation infiltrated) and, how this result compares to a baseline without interventions. Simulation results can be expanded for detail (Figure 4, first row). Through facilitation, participants make sense of the results by focusing on what changed and reason why, based on collective review of the simulation results in successive scenario trials. Such interpretative and explanatory sense-making is enabled by the facilitation and the process-based, simplified model structure; it can be easily conveyed conceptually and made inspectable to participants depending on their comfort level interacting with code—another important reason for using parsimonious models in PM (Zellner 2024).

Figure 4.

Flooding simulation outputs, summarized and expanded.

Another tab shows a group overview of the performance of each scenario (Figure 5). Fora.ai creates graphical scores for each user from L-GrID outputs stored in the SQL database. These scores represent the sum of outputs weighted by their concerns, so that scores can also be compared across users (see Supplemental Materials S4 for more details on these computations). These visual representations allow for further reasoning about intervention impacts on different concerns. The discussion around graphical and numerical scores and spatiotemporal effects feeds back into further intervention designs, but may also lead to reconsideration and adjustment of concerns. If concern profiles are adjusted, scores for all existing scenario trials update immediately. The iteration of this activity helps refine layouts and guide conversation towards collective agreement, including consideration of tradeoffs. Alternatively, participants may decide that their current assumptions are no longer valid, or to reevaluate intervention approaches, or even reformulate the problem itself. All visualizations are saved for each scenario, so that users can easily return to them for review and discussion.

Figure 5.

Group overview showing performance for three hypothetical users across three scenarios; each color corresponds to one output variable contributing to the weighted graphical score.

While we report here on a flooding case study using the L-GrID model, Fora.ai is designed to ultimately embed other models that will perform their own computations. The PM platform is designed to serve as an intuitive user interface that enables users to enter their priorities, rapidly build and simulate intervention scenarios, and to make sense of simulation results through the visualization of spatial outcomes and scores weighted by concern profiles.

Workshop structure, execution, and assessment

Workshops with Fora.ai are structured to meet collaborative learning and solution-building goals co-defined with stakeholders, and to collect data to assess how those goals were met (Zellner 2024). In our study, the workshop was organized around the platform’s interactive whiteboard to collectively design GI solutions to flooding in Chelsea, Massachusetts. The workshops described here were in person, where users: (1) input their individual priorities, (2) collaboratively ran simulations to understand flooding in the neighborhood, (3) co-designed GI scenarios to address flooding impacts, (4) saw how their changes affected the flooding simulation, (5) deliberated on the tradeoffs that arose from each solution due to the group’s diverse priorities, and (6) gradually built a GI strategy as their collective experience informed further rounds of scenario-building and testing and/or the revision of their priorities.

We organized a first workshop with 7 local planners and city staff in the City of Chelsea, Massachusetts, interested in supporting GI planning in the community they serve. A second workshop was held at a conference clinic with 17 earth systems modelers interested in exploring how geophysical models could be made more accessible to stakeholders and decision-makers. In both cases, small groups were organized around a facilitator from our research team or from the participant groups who were briefly trained in advance. In both instances, group composition was fairly homogeneous (all city officials or all scientists), so role playing was suggested to consider other concerns and perspectives.

Ahead of each workshop, participants in our study were provided a tutorial video demonstrating the main features of Fora.ai. The workshops started with a brief introduction and purpose for the activity, followed by a demonstration of the platform that built upon the tutorial on how to interact with the various interface elements to create GI scenarios, run simulations, and visualize the various outputs. Participants were then grouped around their assigned facilitator and were given a group-specific URL to access a Fora.ai breakout room. In their breakout room, they could engage in the exploration of flooding in Chelsea and in the design and testing of GI solutions to address it. A planned neighborhood project was included in the baseline scenario so participants could build on and expand interventions to integrate the project within the fabric of the surrounding neighborhood.

Participants connected to their corresponding Fora.ai breakout room with individual or shared laptops or mobile devices and defined their priorities using the platform’s concern profile. The small groups then engaged in facilitated iterative collaborative GI planning and testing, and reflection on the outcomes of each design, organized around collectively defined success outcomes and systematic exploration strategies (see Supplemental Materials S3 for details on the facilitation structure). Each small group was provided with a flip chart or white board for the facilitator to take notes on goals, scenario strategy (e.g., what to improve and how), effectiveness of the solution relative to the goals, and emerging tradeoffs from each iteration. While the model ran, facilitators recorded participants’ expectations for the outcomes and their rationale on the white board or flip chart. Once the simulations were done, the impacts and tradeoffs were visualized and saved to be discussed relative to the baseline, past scenarios, and expectations, after which the facilitator led the conversation to next strategies based on what was collectively learned. The scenario-building, modeling, and reflection was iterated as many times as time allowed.

The workshop ended with a large-group reflection on insights derived from the experience, the usability of different platform and model features and visualizations, and the overall facilitated process. Participants shared additional thoughts on how they envisioned using Fora.ai for collaborative problem solving and planning, and implications for data collection and interpretation.

Evidence of impact

The scenario and simulation data saved by the platform, the facilitators’ notes and the conversations recorded enabled an in-depth analysis of the collaborative learning and planning afforded by Fora.ai, the model embedded in it, and the facilitation practices. This data provides evidence of impact and helps identify areas for improvement and future developed (see Implications and next steps). We also audio-recorded the small-group and large-group discussions; the recordings with city officials were discarded, however, due to recording malfunction.

For our analysis, we rely on Fora.ai scenarios created by participants’ groups (Figure 6), the simulation outputs weighted by the concern profiles (Figure 7), facilitator and researcher notes, and user quotes. We study the evolution of GI solutions over time, and how simulation results and deliberations supported this change. We qualitatively interpret how the facilitated use of Fora.ai prompted participants to collectively change their understanding of how GI types and layouts mitigate flooding, identify microspatial inequities caused by flooding, surface the tradeoffs from specific GI strategies, and innovate on solutions and design practices. We demonstrate these changes by showcasing examples of conversations and scenarios as groups transitioned from prior beliefs in what would work to new ways of reasoning that supported more impactful and equitable strategies. We also highlight the role of facilitation and of the platform features in supporting these transformations.

Figure 6.

Evolution of GI scenarios and corresponding flooding results produced by Chelsea city officials group 1 (a) and group 2 (b); scientists group 1 (c), group 2 (d) and group 3 (e). Increasing ponding in darker shades of blue; pink = flooding; red squares = bioswales; light blue circles = rain barrels; light green squares = green roofs; slate squares = permeable pavers.

Figure 7.

Simulation results produced by Chelsea city officials group1 (a) and group 2 (b); scientists group 1 (c), group 2 (d) and group 3 (e). Results weighted by the concern profiles of each group member; different colors correspond to the output variable contributing to the weighted graphical score.

Learning about flooding and GI strategies

In both types of users, we observed a growing understanding of how runoff flows and accumulates in the landscape, and of how the amount, types and spatial layouts of GI could influence the outputs for each strategy. This trend was not observed in all small groups, however. We highlight similarities and differences across groups.

Workshop with city officials

Both groups initiated their exploration focusing on a project site planning to include GI. City officials focused their intervention on this site and surrounding block, with green roofs as the GI type of choice due to their many co-benefits (e.g., heat mitigation and pollination), but their first scenario showed little impact (Figures 8 and 9, left). The disappointing results were initially surprising, but the structured facilitation allowed participants to reflect on why. They extended the intervention beyond the project site, focusing on familiar and less costly sites (e.g., a parking lot, as opposed to a building). With further exploration, they realized that bioswales and permeable pavers had more infiltration and storage capacity than other GI types, so they took a more prominent role in their evolving designs. The new layouts were clustered, which limited how much runoff could reach the GI, so capacity used remained low. Realizing this effect, they scattered the GI with more effective results across outputs and at lower costs (Figures 8 and 9, second left). This iterative design process led to gradual increases in graphical scores, but the impacts were limited due to the small amount of GI placed. In the large-group discussion, participants expressed surprise that GI decentralization was so important. Despite the recognized limitations of green roofs and of barrels, participants still highlighted their co-benefits and visibility.

Figure 8.

Simulation results for group 1 of Chelsea city officials. Top: scenario maps. Increasing ponding in darker shades of blue; pink = flooding; red squares = bioswales; light blue circles = rain barrels; light green squares = green roofs; slate squares = permeable pavers. Bottom: scenario score bars. Results weighted by the concern profiles of each group member; different colors correspond to the output variable contributing to the weighted graphical score.

Figure 9.

Simulation results for group 2 of Chelsea city officials. Top: scenario maps. Increasing ponding in darker shades of blue; pink = flooding; red squares = bioswales; light blue circles = rain barrels; light green squares = green roofs; slate squares = permeable pavers. Bottom: scenario score bars. Results weighted by the concern profiles of each group member; different colors correspond to the output variable contributing to the weighted graphical score.

Group 2 tried out more GI types than group 1 (Figure 9), finding that rain barrels and permeable pavers reduced the cost per gallon of water captured relative to green roofs. Their growing understanding of the effectiveness of bioswales along roads informed their final scenario (Figure 9, right), replacing permeable pavers with greater coverage of bioswales. Nevertheless, the group continued with a foundation of GI clustering that, while improving some of the performance metrics, was significantly costlier without a marked benefit.

The platform’s intuitive whiteboard supported the groups’ realizations with minimal assistance from facilitators, but the model outputs needed some clarification. Output visualizations helped participants make sense of the impacts across metrics, and simulation video recordings were useful to understand how water flows caused flooding at various sites. The videos remained separate from the whiteboard, however, which made it hard for users to transfer the information into their designs and sometimes missed the areas they were looking to target.

Facilitation followed the structure described in section 3, but with subtle distinctions. The group 2 facilitator tried to lead the group to reach an effective solution by the end of the workshop, defining and refining success before starting each scenario trial. This organized the discussion but also instilled a sense of urgency—and frustration when solutions were not as impactful as expected. In contrast, the group 1 facilitator focused on increasing understanding about how GI works to address multiple objectives. The latter strategy led to a more systematic exploration of GI types and spatial strategies where users realized the impact of GI decentralization and the limitations of green roofs sooner. Both groups increased their graphical scores, but group 1 derived generalizable principles, while group 2 tried to create more impactful scenarios.

Workshop with scientists

Incorporating lessons learned from the first workshop, we included in the facilitation structure a definition of success at the start of each scenario-building iteration, and a space to reflect on how and why they were met or not.

Figures 10–12, show more varied outcomes than in the workshop with city officials, illustrating how facilitation and group composition can influence the experience. In the workshop with Chelsea city officials, the groups were smaller, and participants knew each other well, which was not always the case for the scientists. The facilitators in this workshop received a brief training, but the role was unfamiliar to them. The facilitator of group 1, however, was an experienced principal investigator of interdisciplinary groups.

Figure 10.

Simulation results for group 1 of scientists. Top: Scenario maps. Increasing ponding in darker shades of blue; pink = flooding; red squares = bioswales; light blue circles = rain barrels; light green squares = green roofs; slate squares = permeable pavers. Bottom: scenario score bars. Results weighted by the concern profiles of each group member; different colors correspond to the output variable contributing to the weighted graphical score.

Figure 11.

Simulation results for group 2 of scientists. Top: scenario maps. Increasing ponding in darker shades of blue; pink = flooding; red squares = bioswales; light blue circles = rain barrels; light green squares = green roofs; slate squares = permeable pavers. Bottom: scenario score bars. Results weighted by the concern profiles of each group member; different colors correspond to the output variable contributing to the weighted graphical score.

Figure 12.

Simulation results for group 3 of scientists. Top: Scenario maps. Increasing ponding in darker shades of blue; pink = flooding; red squares = bioswales; light blue circles = rain barrels; light green squares = green roofs; slate squares = permeable pavers. Bottom: scenario score bars. Results weighted by the concern profiles of each group member; different colors correspond to the output variable contributing to the weighted graphical score.

Group 1 ran a baseline scenario to identify where flooding occurs, as none of the participants was familiar with the area. They used this baseline to guide the design of their first scenario, based on problem areas. They wanted first to keep costs low, but the facilitator prompted them to define success relative to percentage damage or flooding reduction. The group settled for flooding reduction as a goal, with a high-cost strategy to “place everything, everywhere,” prioritizing open spaces (where there was opportunity for GI placement) and areas with consistent flooding, pointing to specific locations on the whiteboard (Figure 10, left). They were happy across users, but the investment skyrocketed (Figure 10, center). The lead researcher then pointed the group to notice the details of some of the output metrics, particularly the GI capacity used, to draw attention to overbuilding. They devised two strategies to pursue: first, create a checker-board pattern of GI to increase surface area for runoff capture; second, prioritize bioswales and pavers over green roofs, for their greater capacity to capture runoff (Figure 10, right). Rain barrels were a low-cost option that they could also add, even with limited capacity. The group thus set a goal for capacity used to include in their definition of success, which satisfied all participants and increased their confidence in their strategy, attaining similar impacts at much lower costs and improving scores across all users (Figure 10, right).

Group 2 started with a funding goal. The lead author stepped in to guide goal setting with various metrics before suggesting GI strategies to meet them. They settled with “protecting lives and property,” or “protecting lives and infrastructure,” which was only partially represented by the output metrics (property damage). The initial proposal gravitated towards green roofs, but aware that this GI type alone could not address flooding. They thus decided on a “multipronged approach, because at some point the water will get on the ground.” After some deliberation over the co-benefits of green roofs, installation versus maintenance costs relative to permeable pavers, and other GI alternatives that the facilitator prompted them to consider, the group decided to set a first scenario solely with green roofs (Figure 11, left). Their spatial strategy was tied to big buildings, which they defined in concrete spatial dimensions as “more than 10 connected pixels,” deemed to be “more efficient.” They found out through the simulation that “Green roofs will do absolutely zero things for flooding…,” but they also noticed that the ponding happened away from residential areas and roads. When they turned to the full simulation results, they realized that damage had not been reduced, even though the capacity of the GI had been fully used. The significant investment this scenario required, and the fact that the valuation of this scenario was very low for some participants prompted them to try permeable surfaces where water was ponding. The location of flooding and ponding (i.e., the microspatial inequities) started to guide their designs.

They turned to permeable pavers, adding rain barrels and bioswales, despite the latter’s expense. They also considered the flow of water and the influence of topography, and engaged in more playful exploration:

Speaker 32: “This is like a free for all wild west type scenario at this point. I love it. It might work. If you let everyone do what they want. Let's see what happens.”

Speaker 33: “…it's like water is going down so we need infiltration.”

They finalized the scenario with bioswales and permeable pavers (Figure 11, center) and reset their goals to reduce flooding and save property, now more closely reflecting the model outputs. They expected that, since their design was “reactive” to where flooding was happening, that it would be more effective in meeting these goals. They were generally pleased to see that all scores had improved, and they had done so lowering the investment costs (Figure 11, center), but half of the GI capacity was used. Building on this spatial strategy, they tried new GI types, clustering large capacity rain barrels to simulate underground storage while decentralizing bioswales to enhance capture. In their final scenario, the facilitator suggested to try rain barrels in one flooded area, permeable pavers in another, to see which one did better. This strategy would allow participants to see the spatial impact on the map, but not on the aggregated metrics. While this trial showed improvement, damage was not reduced (Figure 11, right). Still, they noticed that “the area where we had done permeable pavers that we switched to the swales is now much better” and that “permeable pavers are not helpful.” Rain barrels were used fully, but only because of their limited capacity.

Group 3 had a harder time with this exercise. As they assumed different personas, they prioritized GI co-benefits (rain barrels for gardening, bioswales for urban greening) and abstract goals (sustainability), but not the priorities that were represented in the model. One of the participants was a strong advocate for such valuation, and the facilitator further cemented that approach:

“Okay, cool. …. we're not [going] to be able to measure from a model, but we still value … because we think they're valuable... for other reasons that maybe another model could address.”

This framing led the group to design scenarios that they would not be able to assess with the simulation results. The facilitator inadvertently further constrained the conversation and the willingness of participants to try repeated scenarios, even as he tried to follow the facilitation structure:

“We're gonna probably have like three chances to run different simulations. So, we can … systematically explore this space, that kind of meets some of our questions about the trade offs between the cost and the benefits. It sounds like rain barrels are one because we think that's potentially a sustainable solution. And bioswales are because it seems like a very effective solution.”

The terms “sustainable” or “effective” had not been defined relative to the outputs of the model, which made goals uncertain and unattainable. Rather than engaging in playful learning, participants instead focused on “getting it right” in the few chances they had to succeed. There were moments in which they engaged in “careful placement” of GI in flooded areas or along roads, like other groups, but tried to combine strategies and anticipate the impacts instead of learning through trials to tease out the effects of different strategies. Consequently, the group had a difficult time deciding what to try and why, and ran their first scenario at 40 minutes, relative to the 20 to 25 minutes in other groups. There was no time to reason through the results (Figure 12, left) before starting to design the last scenario, so results were unsatisfactory (Figure 12, right). Looking for the ultimate solution that would “work,” the group moved away from deriving generalizable understanding and designing effective strategies to address flooding relative to the metrics of concern.

In contrast, the other facilitators created a more playful and low-risk environment, encouraging and prompting participants to try things out, illustrated by the group 1 facilitator:

“I have the feeling the second and the third trial will be easier, because we'll see what the fact of the first one is. So I think the first one … will be a lot of money. But we can decide what scenarios we want to try and see what the effect might be. Any suggestions from the consultant or from the mayor of what to do?”

while systematically building on successful outcomes and gradually adjusting scenarios based on their learning:

“[F]or this third scenario, we should try to stick to the same strategy and adjust then, once we know how far off we were. So we can make adjustments.”

Learning about microspatial inequities and how to address them with GI

Participants in both workshops and across all groups increasingly noticed flooding impacts within the neighborhood, but to different extent. We highlight how unequal distribution of effects were identified and influenced GI designs.

Workshop with city officials

In this workshop, the awareness of microspatial inequities was subtle and indirect. They became more visible around the limited performance of opportunistic site-level GI interventions (e.g., on planned projects and parking lots). As participants tested scenarios, they realized that they would not have a widespread effect because flooding was happening in areas that were spatially disconnected to the sites where GI installations were proposed. Discussions expanded to neighborhood-level interventions, but the absence of flooding data on the whiteboard impeded a more explicit exploration of these inequities and how GI could address them. This barrier was mentioned in the large group discussion and informs ongoing development of the platform.

Workshop with scientists

An explicit treatment of microspatial inequity showed up more clearly in the workshop with scientists. After their first scenario trial, groups 1 and 2 focused on the microspatial inequities experienced across the neighborhood (see section 4.1.2), targeting areas with greater flooding for GI placement, which they could identify by replaying the simulation video. Group 2 noticed that their second trial had improved other areas, especially downstream. Strategizing to improve on GI capacity used and reduction in property damage, they strengthened what they called their “reactive” strategy to areas with greatest flooding:

Speaker 33: “I think we should move … to … more targeted flooding areas. Because this is still not working.”

Speaker 30: “I think your thing near Suffolk Street wasn't doing anything.”

Speaker 32: “I think that's where our damage reduction… there's no [flooded area] there anymore.”

The spatial inequities also surfaced as they considered the extent to which decentralization of GI, versus clustering around problematic areas, would be beneficial.

Group 3 did not have enough time to fully consider these spatial inequalities but referred to specific flooded areas as they designed their scenarios.

Negotiating tradeoffs across objectives

Tradeoffs are inherent in complex problems, as effects ripple through the web of system interactions, causing improvements at one scale or dimension at the expense of another. Additionally, diverse stakeholders may value scales and dimensions differently. We identify tradeoff deliberations during GI scenarios.

Workshop with city officials

City officials shared similar concerns, but one participant in group 1 represented the concern profile of state environmental agencies with whom they routinely interacted. In this group, each iteration led to a gradual improvement in scores, as well as more even scores across the profiles (Figure 13(a)). The improvements are not as noticeable as in the workshops with scientists (section 4.3.2); at that time the priorities were not translated into definitions of success to help guide the scenario-building with the tradeoffs in mind. They were, however, for group 2, but the search for a solution—rather than learning—prevented participants from improving on equity across objectives (Figure 13(b)). The lessons learned from this workshop led to more noticeable success in the workshop with scientists.

Figure 13.

Simulation results produced by Chelsea city officials group1 (a) and group 2 (b). Results weighted by the concern profiles of each group member; different colors correspond to the output variable contributing to the weighted graphical score.

Workshop with scientists

The group 1 facilitator initiated the session asking participants to decide their roles and fill out their concern profile accordingly, imagining the experience of different storms from the perspective they represented (e.g., city mayor, resident, and scientist). Participants eased into their roles and shared their profiles, explaining their ranking. Their discovery of emerging tradeoffs across GI types led them to define success as a function of tradeoffs across output variables:

“So I guess we should also define … what is success going to look like? Spending the least amount of money and reduce flood anyway?”

Other participants agreed and added other metrics to represent this tradeoff, like the cost of property damage. Spatial tradeoffs were also considered, as microspatial inequities became apparent, while recognizing that flooding in open areas (e.g., parking lots, parks) may be less damaging than in residential blocks, except along coastal areas due to pollution.

Over subsequent scenarios the conversation centered on who was happy, who wasn’t and why, based on the metrics used to define success: investment and maintenance costs, property damage, and flooded area, occasionally introducing the investment cost per gallon of runoff captured by GI. Towards the end of the experience, the facilitator reflected:

“I think it's just hard to find agreement in those cases. It can get very contentious. There's also different interests... and I live downstream of you, if you change things upstream, it affects me. I think it helps as a scientist to see what people want. To make you aware before you come up with a grand plan.”

While conflicts across interests were not completely resolved within the span of the workshop, the PM process and platform helped visualize tradeoffs and support collaboration across expertise (scientists, public officials, and residents) to address the shared problem. Perhaps because the facilitator ensured that the concern profiles took a prominent role from the start and participants shared them, this group not only improved their solutions in each iteration, but also worked towards more equitable ones across the values expressed through the concern profiles (Figure 14(a)). Importantly, participants did not need to agree on values (represented as different color schemes in the score bars) to agree on solutions. The reflection on the role of scientists is particularly notable, acknowledging that scientific solutions are not universally acceptable or relevant to the social context, suggesting that they cannot be imposed over other value systems.

Figure 14.

Simulation results produced by scientists group 1 (a), group 2 (b), and group 3 (c). Results weighted by the concern profiles of each group member; different colors correspond to the output variable contributing to the weighted graphical score.

Group 2 expressed initial hesitancy about representing different value sets but engaged with their roles. In the initial scenario-building exercise, the business representative highlighted the aesthetic and the job-generating appeal of green roofs to promote that strategy, while others reflected on the expense of the various GI types. As they tried their first scenario with green roofs on big buildings, they explored the impacts on happiness across perspectives. Given the disparities, the facilitator asked the group how to make everyone happier, and “optimize” across outputs of interest. GI capacity was consistently underutilized and became a focus to also help reduce overbuilding, while targeting flooded buildings to improve on property damage and being mindful of the cost of each GI type. In their final scenario, they realized that capacity used, by itself, may be a poor measure of performance. This was illustrated by the result that rain barrels easily achieved 100% due to their limited capacity, but their impact on all other metrics was negligible. Their scores improved overall, but the tradeoffs across them remained (Figure 14(b)).

Group 3 cared deeply about cost, but without tradeoffs with other model metrics. Their second scenario was named “cheap and fast,” but did not improve other outputs due to the difficulties described above. Nevertheless, this group realized how costly GI interventions may be, and reflected on the need for a systematic exploration of the solution space. While some scores improved for some users, many tradeoffs were unaddressed (Figure 14(c)).

Innovating GI design and planning practices

One challenge PM seeks to address is the fixation on popular solutions that may work well for one dimension or at one scale—typically the site—but not across scales or priorities (Zellner 2024). Another challenge is the expert-driven design process with limited public participation, which tends to be contentious. In such institutional context, there is little room for knowledge co-generation to inform plans and examine their effectiveness across diverse values and scales. We expected that, through collaborative design and modeling, workshop participants would move away from preconceived attachments and develop more concrete, nuanced, and equitable GI strategies based on their growing understanding of flooding dynamics, microspatial inequities, and the tradeoffs emerging from proposed solutions. Participants realized that their initial favored solutions (e.g., green roofs, site-level or clustered interventions) had limited impact, reasoned through why that was, and shifted their exploration to alternative GI designs. These were increasingly tailored to the problem and the diversity of values represented and attended to both localized flooding and neighborhood-level impacts. Participants also derived design principles that could be applied elsewhere. We summarize here the transitions as they reasoned through small group deliberation and their final large-group reflection.

Workshop with city officials

City officials were surprised that decentralizing GI was more effective compared to concentrated layouts but realized that scattered pattern had greater chances of capturing stormwater runoff. Relatedly, they noted that some locations flooded persistently, regardless of GI spatial strategy; their initial scenarios focused on small interventions that would not scale up to neighborhood-level impacts. Regarding GI type, participants admitted that green roofs were expensive and not cost-effective for flooding but would still advocate for them due to their co-benefits. One of the groups gained a better sense of how permeable pavers could outperform green roofs at relatively lower costs for greater capacity, but pavers also needed to be strategically placed along specific roads to ensure stormwater would reach it. While rain barrels would not do much to reduce flooding (even with larger capacities), participants highlighted their political co-benefits, making GI efforts visible. On planning practice, participants were eager to continue interacting with the platform to further inform GI planning and engage others in the design and deliberative process.

Workshop with scientists

Earth system modelers may be more comfortable modeling flooding, but typically on larger and less heterogeneous landscapes. Further, GI performance or social concerns are rarely represented. Scientists were also surprised by the limited capacity of green roofs to hold stormwater runoff and reflected on how a greater GI coverage—and investment—was necessary to make a noticeable impact. There was also a growing understanding that strategic decentralization following water flow paths and accumulation would have the greatest impact across metrics while lowering costs. The role of the scientific expert, and how science alone cannot provide solutions to socio-environmental problems was a notable insight. Making the spatial (upstream decisions and downstream effects) and social trade-offs visible prompted the reflection that the scientists’ “grand plan” may not be effective for such complex—and potentially contentious—problems. Some participants commented on how common practice is to build and run simulation models on their own rather than collaboratively; as scientists, they had little exposure to scenario-building deliberations, whether “value-based” (considering priorities) or “information-based” (considering where flooding happens). Like with city officials, they expressed interest in continuing explorations with the platform and tools, trying “a bunch of different things, see how it pans out.”

Implications and next steps

We seek to inform the development and use of DT in ways that can impactfully support environmental planning and policy, strengthening the data and simulation sense-making, and integrating social dimensions. To sustainably and equitably address complex socio-environmental problems, digital tools should enable a coordinated collaborative approach to decision-making. We developed and tested a PM platform, Fora.ai, to examine how it could deliver on some of these goals. Our analysis of the modeling outputs, GI designs, and deliberations shows a transformative effect in each workshop and in most groups. The transition from a priori favored solutions to collaboratively novel ones; the increased specificity and overall performance in each successive design; the greater clarity on what worked, where and why towards generalizable insights; the increased equity of solutions across diverse values, and the reflections on current and modeling, design and planning practices, all point to the impactful potential of collaborative solution-building and planning platforms like Fora.ai, but also to the unique facilitation structures and skills that need to support such outcomes. Our experience thus illustrates the advantages of considering a broader PM and decision-making framework to integrate the social dimension into the design of DT for urban planning and management, beyond the focus on technological advances alone. We propose that such approaches and tools are present as early as possible in planning processes, and over time, become seamlessly integrated as an ongoing practice of relationship building and collaborative planning. Our past and current community collaborations have shown the benefits of introducing Fora.ai to inform concrete projects and more general strategies at any decision-making stage of the planning process.

City officials and scientists enjoyed the experience and expressed interest in continuing explorations with improved versions of Fora.ai. While output metrics where unclear for some, the whiteboard was intuitive to navigate. During the activity, participants repeatedly pointed to areas of the map to guide scenario-building and actively helped each other as they learned to use the platform features. One participant used their device to show the baseline simulation as others referred to it to place GI on their devices, further cementing the collaborative nature of the experience.

Yet, we are mindful of limitations to address. One is the difficulty imposed by solutionism, that is, the use of Fora.ai as a producer of an “optimal” or “right” solution. In the large-group discussion, scientists reflected on the open-endedness of the process: they did not know how to approach a solution, and the lack of a flooding data layer made it hard for them to decide on where to intervene. This difficulty was reduced by the ability to conduct multiple scenarios, where each new one was easier, “as long as people don’t go rogue.” This highlights the facilitation challenge, where scientists prefer running simulations individually and then “pitch an idea to the group” over working collaboratively. Conventional practices are not conducive to surfacing and resolving conflicts, where social dimensions become critical to the solution-building process. Facilitation was most helpful when it steered the conversation away from the pursuit of accurate predictions and definitive solutions, and instead focused on systematic but playful exploration of spatial and infrastructural strategies, reflection and reasoning through the results across priorities, and redefinition of success relative to the model outputs. This practice more readily produced generalizable strategies and concrete implementations that increased overall performance and equity. Furthermore, participants devised a deliberate approach to GI design and testing; more important than achieving the ultimate solution was developing a way to collaboratively figure it out, strengthening collective reasoning ability and social capital along the way. This serves as an initial step towards outlining a general strategy that the participants could endorse, that would ideally be followed with more detailed engineering and landscape design.

Technological advancements can ensure agile, transparent, and intuitive modeling and visualizations, adaptive to the needs and capabilities of diverse users and facilitators, thus supporting collaborative exploration, sense-making and synthesis for policy and planning. Participants helped us identify several interface improvements to make. Viewing the aggregate scores was useful for overall assessment of scenarios and goal setting, but spatialized—and less pixelated—data layers would have helped guide the subsequent designs to achieve the goals set and are currently being tested. The investment score also caused confusion, and reformulation is underway to include a penalty on the whole score with increasing spending over the established budget. L-GrID model extensions are underway to help strengthen the deliberation of tradeoffs across a broader range of impacts, GI ecosystem functions, and areas. Our participants also envisioned a broad range of applications for Fora.ai (e.g., watershed infrastructure siting, blue infrastructure planning, wildfire management), afforded by its ability to plug in different models and engage diverse participants, as proposed by Ferré-Bigorra et al. (2022). While more realistic 3D representations may be warranted for some of these applications, humans, rather than the technology, remain at the center of such socio-technical system.

In line with DT technology, Fora.ai and its embedded models—current and planned—are designed to represent hyper-local conditions that enable concrete and tailored design. Where data were not available in our study, they were explicitly assumed or simulated, making the parameters and outputs inspectable. The PM process with Fora.ai not only provides a collaborative sense-making framework for model parameterization and validation but also for sensor network design and data interpretation to feed back into further modeling and exploration. While not deployed at this stage of our project, we anticipate that the addition of sensor data to Fora.ai will greatly enhance the trust in the insights produced through PM and resolve the ambiguities inherent in complex systems modeling due to either structural uncertainty or social diversity. Work is currently ongoing to operationalize the meaningful two-way information flows—even if not in real time—aspired in DT. As proposed, collaborative models, platforms, and facilitation approaches generate the collective critical thinking, complex systems literacy, and social capital needed to make sense of simulated and sensor data to effectively inform planning and policy.

Supplemental Material

Supplemental Material - Enhancing digital twin technology with community-led, science-driven participatory modeling: A case in green infrastructure planning

Supplemental Material for Enhancing digital twin technology with community-led, science-driven participatory modeling: A case in green infrastructure planning by Moira L Zellner, Dean Massey, Michelle Laboy, Daniel T O’Brien, Amy Mueller and Daniel Engelberg in Environment and Planning B: Urban Analytics and City Science.